Skip to content
GitLab
Commit 536ff477 authored by Vivia Nikolaidou's avatar Vivia Nikolaidou 🦆
Browse files

deinterlace: Add yadif ASM optimisations 

Measured to be about 3.4x faster than C

Part-of: <gstreamer/gst-plugins-good!621>
parent ef78014d
Pipeline #161470 waiting for manual action with stages
in 28 seconds
Expand all Hide whitespace changes
Inline Side-by-side
gst/deinterlace/meson.build
View file @ 536ff477
......@@ -33,8 +33,60 @@ else
copy : true)
endif
asm_gen_objs = []
if have_nasm
if host_system == 'windows'
outputname = '@PLAINNAME@.obj'
else
outputname = '@PLAINNAME@.o'
endif
if get_option('b_staticpic')
asm_pic_def = '-DPIC'
else
asm_pic_def = '-UPIC'
endif
# Assembly has to be told when the symbols have to be prefixed with _
if cc.symbols_have_underscore_prefix()
asm_prefix_def = '-DPREFIX'
else
asm_prefix_def = '-UPREFIX'
endif
asm_arch_def = '-DARCH_X86_64=1'
if host_system == 'windows'
asm_outformat = 'win64'
elif ['darwin', 'ios'].contains(host_system)
asm_outformat = 'macho64'
elif host_system.endswith('bsd')
asm_outformat = 'aoutb'
else
asm_outformat = 'elf64'
endif
asm_x = files('x86/yadif.asm',
'x86/x86inc.asm')
asm_stackalign_def = '-DSTACK_ALIGNMENT=64'
asm_incdir = 'x86'
message('Nasm configured on x86-64')
asm_gen = generator(nasm,
output: outputname,
arguments: ['-I@CURRENT_SOURCE_DIR@',
'-I@CURRENT_SOURCE_DIR@/@0@/'.format(asm_incdir),
asm_arch_def,
asm_stackalign_def,
asm_pic_def,
asm_prefix_def,
'-f', asm_outformat,
'-o', '@OUTPUT@',
'@INPUT@'])
asm_gen_objs = asm_gen.process(asm_x)
endif
gstdeinterlace = library('gstdeinterlace',
interlace_sources, orc_c, orc_h,
interlace_sources, asm_gen_objs, orc_c, orc_h,
c_args : gst_plugins_good_args,
include_directories : [configinc],
dependencies : [orc_dep, gstbase_dep, gstvideo_dep],
......
gst/deinterlace/x86/x86inc.asm 0 → 100644
View file @ 536ff477
This diff is collapsed.
gst/deinterlace/x86/yadif.asm 0 → 100644
View file @ 536ff477
;*****************************************************************************
;* x86-optimized functions for yadif filter
;* Copyright (C) 2020 Vivia Nikolaidou <vivia.nikolaidou@ltnglobal.com>
;*
;* Based on libav's vf_yadif.asm file
;* Copyright (C) 2006 Michael Niedermayer <michaelni@gmx.at>
;* Copyright (c) 2013 Daniel Kang <daniel.d.kang@gmail.com>
;*
;* This file is part of FFmpeg.
;*
;* FFmpeg is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* FFmpeg is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with FFmpeg; if not, write to the Free Software
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;******************************************************************************
%include "x86inc.asm"
SECTION_RODATA
; 16 bytes of value 1
pb_1: times 16 db 1
; 8 words of value 1
pw_1: times 8 dw 1
SECTION .text
%macro ABS1 2
%if cpuflag(ssse3)
pabsw %1, %1
%elif cpuflag(mmxext) ; a, tmp
pxor %2, %2
psubw %2, %1
pmaxsw %1, %2
%else ; a, tmp
pxor %2, %2
pcmpgtw %2, %1
pxor %1, %2
psubw %1, %2
%endif
%endmacro
%macro CHECK 2
; %1 = 1+j, %2 = 1-j
; m2 = t0[x+1+j]
movu m2, [tzeroq+%1]
; m3 = b0[x+1-j]
movu m3, [bzeroq+%2]
; m4 = t0[x+1+j]
mova m4, m2
; m5 = t0[x+1+j]
mova m5, m2
; m4 = xor(t0[x+1+j], b0[x+1-j]
pxor m4, m3
pavgb m5, m3
; round down to 0
pand m4, [pb_1]
; m5 = rounded down average of the whole thing
psubusb m5, m4
; shift by 1 quadword to prepare for spatial_pred
psrldq m5, 1
; m7 = 0
; Interleave low-order bytes with 0
; so one pixel doesn't spill into the next one
punpcklbw m5, m7
; m4 = t0[x+1+j] (reset)
mova m4, m2
; m2 = t0[x+1+j] - b0[x+1-j]
psubusb m2, m3
; m3 = -m2
psubusb m3, m4
; m2 = FFABS(t0[x+1+j] - b0[x+1-j]);
pmaxub m2, m3
; m3 = FFABS(t0[x+1+j] - b0[x+1-j]);
mova m3, m2
; m4 = FFABS(FFABS(t0[x+1+j] - b0[x+1-j]);
mova m4, m2
; m3 = FFABS(t0[x+j] - b0[x-j])
psrldq m3, 1
; m4 = FFABS(t0[x-1+j] - b0[x-1-j])
psrldq m4, 2
; prevent pixel spilling for all of them
punpcklbw m2, m7
punpcklbw m3, m7
punpcklbw m4, m7
paddw m2, m3
; m2 = score
paddw m2, m4
%endmacro
%macro CHECK1 0
; m0 was spatial_score
; m1 was spatial_pred
mova m3, m0
; compare for greater than
; each word will be 1111 or 0000
pcmpgtw m3, m2
; if (score < spatial_score) spatial_score = score;
pminsw m0, m2
; m6 = the mask
mova m6, m3
; m5 = becomes 0 if it should change
pand m5, m3
; nand: m3 = becomes 0 if it should not change
pandn m3, m1
; m3 = put them together in an OR
por m3, m5
; and put it in spatial_pred
mova m1, m3
%endmacro
%macro CHECK2 0
; m6 was the mask from CHECK1 (we don't change it)
paddw m6, [pw_1]
; shift words left while shifting in 14 0s (16 - j)
; essentially to not recalculate the mask!
psllw m6, 14
; add it to score
paddsw m2, m6
; same as CHECK1
mova m3, m0
pcmpgtw m3, m2
pminsw m0, m2
pand m5, m3
pandn m3, m1
por m3, m5
mova m1, m3
%endmacro
%macro LOAD 2
movh %1, %2
punpcklbw %1, m7
%endmacro
%macro FILTER_HEAD 0
; m7 = 0
pxor m7, m7
; m0 = c
LOAD m0, [tzeroq]
; m1 = e
LOAD m1, [bzeroq]
; m3 = mp
LOAD m3, [mpq]
; m2 = m1
LOAD m2, [moneq]
; m4 = mp
mova m4, m3
; m3 = m1 + mp
paddw m3, m2
; m3 = d
psraw m3, 1
; rsp + 0 = d
mova [rsp+ 0], m3
; m2 = m1 - mp
psubw m2, m4
; m2 = temporal_diff0 (m4 is temporary)
ABS1 m2, m4
; m3 = t2
LOAD m3, [ttwoq]
; m4 = b2
LOAD m4, [btwoq]
; m3 = t2 - c
psubw m3, m0
; m4 = b2 - e
psubw m4, m1
; m3 = ABS(t2 - c)
ABS1 m3, m5
; m4 = ABS(b2 - e)
ABS1 m4, m5
paddw m3, m4
psrlw m2, 1
; m3 = temporal_diff1
psrlw m3, 1
; m2 = left part of diff
pmaxsw m2, m3
; m3 = tp2
LOAD m3, [tptwoq]
; m4 = bp2
LOAD m4, [bptwoq]
psubw m3, m0
psubw m4, m1
ABS1 m3, m5
ABS1 m4, m5
paddw m3, m4
; m3 = temporal_diff2
psrlw m3, 1
; m2 = diff (for real)
pmaxsw m2, m3
; rsp + 16 = diff
mova [rsp+16], m2
; m1 = e + c
paddw m1, m0
; m0 = 2c
paddw m0, m0
; m0 = c - e
psubw m0, m1
; m1 = spatial_pred
psrlw m1, 1
; m0 = FFABS(c-e)
ABS1 m0, m2
; m2 = t0[x-1]
; if it's unpacked it should contain 4 bytes
movu m2, [tzeroq-1]
; m3 = b0[x-1]
movu m3, [bzeroq-1]
; m4 = t0[x-1]
mova m4, m2
; m2 = t0[x-1]-b0[x-1] unsigned packed
psubusb m2, m3
; m3 = m3 - m4 = b0[x-1]-t0[x-1] = -m2 unsigned packed
psubusb m3, m4
; m2 = max(m2, -m2) = abs(t0[x-1]-b0[x-1])
pmaxub m2, m3
%if mmsize == 16
; m3 = m2 >> 2quadwords
; pixel jump: go from x-1 to x+1
mova m3, m2
psrldq m3, 2
%else
pshufw m3, m2, q0021
%endif
; m7 = 0
; unpack and interleave low-order bytes
; to prevent pixel spilling when adding
punpcklbw m2, m7
punpcklbw m3, m7
paddw m0, m2
paddw m0, m3
; m0 = spatial_score
psubw m0, [pw_1]
CHECK -2, 0
CHECK1
CHECK -3, 1
CHECK2
CHECK 0, -2
CHECK1
CHECK 1, -3
CHECK2
; now m0 = spatial_score, m1 = spatial_pred
; m6 = diff
mova m6, [rsp+16]
%endmacro
%macro FILTER_TAIL 0
; m2 = d
mova m2, [rsp]
; m3 = d
mova m3, m2
; m2 = d - diff
psubw m2, m6
; m3 = d + diff
paddw m3, m6
; m1 = max(spatial_pred, d-diff)
pmaxsw m1, m2
; m1 = min(d + diff, max(spatial_pred, d-diff))
; m1 = spatial_pred
pminsw m1, m3
; Converts 8 signed word integers into 16 unsigned byte integers with saturation
packuswb m1, m1
; dst = spatial_pred
movh [dstq], m1
; half the register size
add dstq, mmsize/2
add tzeroq, mmsize/2
add bzeroq, mmsize/2
add moneq, mmsize/2
add mpq, mmsize/2
add ttwoq, mmsize/2
add btwoq, mmsize/2
add tptwoq, mmsize/2
add bptwoq, mmsize/2
add ttoneq, mmsize/2
add ttpq, mmsize/2
add bboneq, mmsize/2
add bbpq, mmsize/2
%endmacro
%macro FILTER_MODE0 0
.loop0:
FILTER_HEAD
; m2 = tt1
LOAD m2, [ttoneq]
; m4 = ttp
LOAD m4, [ttpq]
; m3 = bb1
LOAD m3, [bboneq]
; m5 = bbp
LOAD m5, [bbpq]
paddw m2, m4
paddw m3, m5
; m2 = b
psrlw m2, 1
; m3 = f
psrlw m3, 1
; m4 = c
LOAD m4, [tzeroq]
; m5 = d
mova m5, [rsp]
; m7 = e
LOAD m7, [bzeroq]
; m2 = b - c
psubw m2, m4
; m3 = f - e
psubw m3, m7
; m0 = d
mova m0, m5
; m5 = d - c
psubw m5, m4
; m0 = d - e
psubw m0, m7
; m4 = b - c
mova m4, m2
; m2 = FFMIN(b-c, f-e)
pminsw m2, m3
; m3 = FFMAX(f-e, b-c)
pmaxsw m3, m4
; m2 = FFMAX(d-c, FFMIN(b-c, f-e))
pmaxsw m2, m5
; m3 = FFMIN(d-c, FFMAX(f-e, b-c))
pminsw m3, m5
; m2 = max
pmaxsw m2, m0
; m3 = min
pminsw m3, m0
; m4 = 0
pxor m4, m4
; m6 = MAX(diff, min)
pmaxsw m6, m3
; m4 = -max
psubw m4, m2
; m6 = diff
pmaxsw m6, m4
FILTER_TAIL
; r13m = w
sub DWORD r13m, mmsize/2
jg .loop0
%endmacro
%macro FILTER_MODE2 0
.loop2:
FILTER_HEAD
FILTER_TAIL
; r13m = w
sub DWORD r13m, mmsize/2
jg .loop2
%endmacro
%macro YADIF_ADD3 0
; start 3 pixels later
add dstq, 3
add tzeroq, 3
add bzeroq, 3
add moneq, 3
add mpq, 3
add ttwoq, 3
add btwoq, 3
add tptwoq, 3
add bptwoq, 3
add ttoneq, 3
add ttpq, 3
add bboneq, 3
add bbpq, 3
%endmacro
; cglobal foo, 2,3,7,0x40, dst, src, tmp
; declares a function (foo) that automatically loads two arguments (dst and
; src) into registers, uses one additional register (tmp) plus 7 vector
; registers (m0-m6) and allocates 0x40 bytes of stack space.
%macro YADIF_MODE0 0
cglobal yadif_filter_line_mode0, 13, 14, 8, 80, dst, tzero, bzero, mone, mp, \
ttwo, btwo, tptwo, bptwo, ttone, \
ttp, bbone, bbp, w
YADIF_ADD3
FILTER_MODE0
RET
%endmacro
%macro YADIF_MODE2 0
cglobal yadif_filter_line_mode2, 13, 14, 8, 80, dst, tzero, bzero, mone, mp, \
ttwo, btwo, tptwo, bptwo, ttone, \
ttp, bbone, bbp, w
YADIF_ADD3
FILTER_MODE2
RET
%endmacro
; declares two functions for ssse3, and two for sse2
INIT_XMM ssse3
YADIF_MODE0
YADIF_MODE2
INIT_XMM sse2
YADIF_MODE0
YADIF_MODE2
gst/deinterlace/yadif.c
View file @ 536ff477
......@@ -31,6 +31,7 @@
#include <gst/gst.h>
#ifdef HAVE_ORC
#include <orc/orc.h>
#include <orc/orcsse.h>
#endif
#include "gstdeinterlacemethod.h"
#include "yadif.h"
......@@ -86,6 +87,41 @@ static void
filter_scanline_yadif_packed_3 (GstDeinterlaceSimpleMethod * self,
guint8 * out, const GstDeinterlaceScanlineData * scanlines, guint size);
static void
filter_line_c_planar_mode0 (void *ORC_RESTRICT dst,
const void *ORC_RESTRICT tzero, const void *ORC_RESTRICT bzero,
const void *ORC_RESTRICT mone, const void *ORC_RESTRICT mp,
const void *ORC_RESTRICT ttwo, const void *ORC_RESTRICT btwo,
const void *ORC_RESTRICT tptwo, const void *ORC_RESTRICT bptwo,
const void *ORC_RESTRICT ttone, const void *ORC_RESTRICT ttp,
const void *ORC_RESTRICT bbone, const void *ORC_RESTRICT bbp, int w);
static void
filter_line_c_planar_mode2 (void *ORC_RESTRICT dst,
const void *ORC_RESTRICT tzero, const void *ORC_RESTRICT bzero,
const void *ORC_RESTRICT mone, const void *ORC_RESTRICT mp,
const void *ORC_RESTRICT ttwo, const void *ORC_RESTRICT btwo,
const void *ORC_RESTRICT tptwo, const void *ORC_RESTRICT bptwo,
const void *ORC_RESTRICT ttone, const void *ORC_RESTRICT ttp,
const void *ORC_RESTRICT bbone, const void *ORC_RESTRICT bbp, int w);
static void (*filter_mode2) (void *ORC_RESTRICT dst,
const void *ORC_RESTRICT tzero, const void *ORC_RESTRICT bzero,
const void *ORC_RESTRICT mone, const void *ORC_RESTRICT mp,
const void *ORC_RESTRICT ttwo, const void *ORC_RESTRICT btwo,
const void *ORC_RESTRICT tptwo, const void *ORC_RESTRICT bptwo,
const void *ORC_RESTRICT ttone, const void *ORC_RESTRICT ttp,
const void *ORC_RESTRICT bbone, const void *ORC_RESTRICT bbp, int w);
static void (*filter_mode0) (void *ORC_RESTRICT dst,
const void *ORC_RESTRICT tzero, const void *ORC_RESTRICT bzero,
const void *ORC_RESTRICT mone, const void *ORC_RESTRICT mp,
const void *ORC_RESTRICT ttwo, const void *ORC_RESTRICT btwo,
const void *ORC_RESTRICT tptwo, const void *ORC_RESTRICT bptwo,
const void *ORC_RESTRICT ttone, const void *ORC_RESTRICT ttp,
const void *ORC_RESTRICT bbone, const void *ORC_RESTRICT bbp, int w);
static void
copy_scanline (GstDeinterlaceSimpleMethod * self, guint8 * out,
const GstDeinterlaceScanlineData * scanlines, guint size)
......@@ -139,36 +175,31 @@ static void
dism_class->interpolate_scanline_nv21 = filter_scanline_yadif_semiplanar;
}
static void
gst_deinterlace_method_yadif_init (GstDeinterlaceMethodYadif * self)
{
}
#define FFABS(a) ABS(a)
#define FFMIN(a,b) MIN(a,b)
#define FFMAX(a,b) MAX(a,b)
#define FFMAX3(a,b,c) FFMAX(FFMAX(a,b),c)
#define FFMIN3(a,b,c) FFMIN(FFMIN(a,b),c)
#define CHECK(j)\
{ int score = FFABS(s->t0[x - colors2 * (1 + (j))] - s->b0[x - colors2 * (1 - (j))])\
+ FFABS(s->t0[x + colors2 * (j)] - s->b0[x -colors2 * (j)])\
+ FFABS(s->t0[x + colors2 * (1 + (j))] - s->b0[x + colors2 * (1 - (j))]);\
#define CHECK(j1, j2, j3)\
{ int score = FFABS(stzero[x - j1] - sbzero[x - j2])\
+ FFABS(stzero[x + j3] - sbzero[x - j3])\
+ FFABS(stzero[x + j1] - sbzero[x + j2]);\
if (score < spatial_score) {\
spatial_score= score;\
spatial_pred= (s->t0[x +colors2 * ((j))] + s->b0[x - colors2 * (j)])>>1;\
spatial_pred= (stzero[x + j3] + sbzero[x - j3])>>1;\
/* The is_not_edge argument here controls when the code will enter a branch
* which reads up to and including x-3 and x+3. */
#define FILTER(start, end, is_not_edge) \
for (x = start; x < end; x++) { \
int c = s->t0[x]; \
int d = (s->m1[x] + s->mp[x])>>1; \
int e = s->b0[x]; \
int temporal_diff0 = FFABS(s->m1[x] - s->mp[x]); \
int temporal_diff1 =(FFABS(s->t2[x] - c) + FFABS(s->b2[x] - e) )>>1; \
int temporal_diff2 =(FFABS(s->tp2[x] - c) + FFABS(s->bp2[x] - e) )>>1; \
int c = stzero[x]; \
int d = (smone[x] + smp[x])>>1; \
int e = sbzero[x]; \
int temporal_diff0 = FFABS(smone[x] - smp[x]); \
int temporal_diff1 =(FFABS(sttwo[x] - c) + FFABS(sbtwo[x] - e) )>>1; \
int temporal_diff2 =(FFABS(stptwo[x] - c) + FFABS(sbptwo[x] - e) )>>1; \
int diff = FFMAX3(temporal_diff0 >> 1, temporal_diff1, temporal_diff2); \
int spatial_pred = (c+e) >> 1; \
int colors2 = colors; \
......@@ -177,15 +208,21 @@ gst_deinterlace_method_yadif_init (GstDeinterlaceMethodYadif * self)
colors2 = 2; \
\
if (is_not_edge) {\
int spatial_score = FFABS(s->t0[x-colors2] - s->b0[x-colors2]) + FFABS(c-e) \
+ FFABS(s->t0[x+colors2] - s->b0[x+colors2]); \
CHECK(-1) CHECK(-2) }} }} \
CHECK( 1) CHECK( 2) }} }} \
int spatial_score = FFABS(stzero[x-colors2] - sbzero[x-colors2]) + FFABS(c-e) \
+ FFABS(stzero[x+colors2] - sbzero[x+colors2]); \
int twice_colors2 = colors2 << 1; \
int minus_colors2 = -colors2; \
int thrice_colors2 = colors2 * 3; \
int minus2_colors2 = colors2 * -2; \
CHECK(0, twice_colors2, minus_colors2) \
CHECK(-colors2, thrice_colors2, minus2_colors2) }} }} \
CHECK(twice_colors2, 0, colors2) \
CHECK(thrice_colors2, minus_colors2, twice_colors2) }} }} \
}\
\
if (!(mode&2)) { \
int b = (s->tt1[x] + s->ttp[x])>>1; \
int f = (s->bb1[x] + s->bbp[x])>>1; \
int b = (sttone[x] + sttp[x])>>1; \
int f = (sbbone[x] + sbbp[x])>>1; \
int max = FFMAX3(d - e, d - c, FFMIN(b - c, f - e)); \
int min = FFMIN3(d - e, d - c, FFMAX(b - c, f - e)); \
\
......@@ -197,16 +234,20 @@ gst_deinterlace_method_yadif_init (GstDeinterlaceMethodYadif * self)
else if (spatial_pred < d - diff) \
spatial_pred = d - diff; \
\
dst[x] = spatial_pred; \
sdst[x] = spatial_pred; \
\
}
ALWAYS_INLINE static void
filter_line_c (guint8 * dst,
const GstDeinterlaceScanlineData * s, int start, int end, int mode,
int colors, int y_alternates_every)
filter_line_c (guint8 * sdst, const guint8 * stzero, const guint8 * sbzero,
const guint8 * smone, const guint8 * smp, const guint8 * sttwo,
const guint8 * sbtwo, const guint8 * stptwo, const guint8 * sbptwo,
const guint8 * sttone, const guint8 * sttp, const guint8 * sbbone,
const guint8 * sbbp, int w, int colors, int y_alternates_every, int start,
int end, int mode)
{
int x;
/* The function is called for processing the middle
* pixels of each line, excluding 3 at each end.
* This allows the FILTER macro to be
......@@ -218,9 +259,74 @@ filter_line_c (guint8 * dst,